mbox series

[v4,00/40] lib/find: add atomic find_bit() primitives

Message ID 20240620175703.605111-1-yury.norov@gmail.com (mailing list archive)
Headers show
Series lib/find: add atomic find_bit() primitives | expand

Message

Yury Norov June 20, 2024, 5:56 p.m. UTC
--- 

This v4 moves new API to separate headers, as adding stuff to find.h
concerns people, particularly Linus. It also adds few more conversions
alongside other cosmetic changes. See full changelog below.

---

Add helpers around test_and_{set,clear}_bit() to allow searching for
clear or set bits and flipping them atomically.

Using atomic search primitives allows to implement lockless bitmap
handling where only individual bits are touched by concurrent processes,
and where people now have to protect their bitmaps to search for a free
or set bit due to the lack of atomic searching routines.

The typical lock-protected bit allocation may look like this:

	unsigned long alloc_bit()
	{
		unsigned long bit;

		spin_lock(bitmap_lock);
		bit = find_first_zero_bit(bitmap, nbits);
		if (bit < nbits)
			__set_bit(bit, bitmap);
		spin_unlock(bitmap_lock);

		return bit;
	}

	void free_bit(unsigned long bit)
	{
		spin_lock(bitmap_lock);
		__clear_bit(bit, bitmap);
		spin_unlock(bitmap_lock);
	}

Now with atomic find_and_set_bit(), the above can be implemented
lockless, directly by using it and atomic clear_bit().

Patches 36-40 do this in few places in the kernel where the
transition is clear. There is likely more candidates for
refactoring.

The other important case is when people opencode atomic search
or atomic traverse on the maps with the patterns looking like:

	for (idx = 0; idx < nbits; idx++)
		if (test_and_clear_bit(idx, bitmap))
			do_something(idx);

Or like this:

	do {
		bit = find_first_bit(bitmap, nbits);
		if (bit >= nbits)
			return nbits;

	} while (!test_and_clear_bit(bit, bitmap));

	return bit;

In both cases, the opencoded loop may be converted to a single function
or iterator call. Correspondingly:

	for_each_test_and_clear_bit(idx, bitmap, nbits)
		do_something(idx);

Or:
	return find_and_clear_bit(bitmap, nbits);

Obviously, the less routine code people have to write themself, the
less probability to make a mistake. The patch #33 fixes one such
mistake.

The new API is not only a handy helpers - it also resolves a non-trivial
issue of using non-atomic find_bit() together with atomic
test_and_{set,clear)_bit().

The trick is that find_bit() implies that the bitmap is a regular
non-volatile piece of memory, and compiler is allowed to use such
optimization techniques like re-fetching memory instead of caching it.

For example, find_first_bit() is implemented like:

      for (idx = 0; idx * BITS_PER_LONG < sz; idx++) {
              val = addr[idx];
              if (val) {
                      sz = min(idx * BITS_PER_LONG + __ffs(val), sz);
                      break;
              }
      }

On register-memory architectures, like x86, compiler may decide to
access memory twice - first time to compare against 0, and second time
to fetch its value to pass it to __ffs().

When running find_first_bit() on volatile memory, the memory may get
changed in-between, and for instance, it may lead to passing 0 to
__ffs(), which is an undefined behaviour. This is a potentially
dangerous call.

find_and_clear_bit() as a wrapper around test_and_clear_bit()
naturally treats underlying bitmap as a volatile memory and prevents
compiler from such optimizations.

Now that KCSAN is catching exactly this type of situations and warns on
undercover memory modifications. We can use it to reveal improper usage
of find_bit(), and convert it to atomic find_and_*_bit() as appropriate.

In some cases concurrent operations with plain find_bit() are acceptable.
For example:

 - two threads running find_*_bit(): safe wrt ffs(0) and returns correct
   value, because underlying bitmap is unchanged;
 - find_next_bit() in parallel with set or clear_bit(), when modifying
   a bit prior to the start bit to search: safe and correct;
 - find_first_bit() in parallel with set_bit(): safe, but may return wrong
   bit number;
 - find_first_zero_bit() in parallel with clear_bit(): same as above.

In last 2 cases find_bit() may not return a correct bit number, but
it may be OK if caller requires any (not exactly the first) set or clear
bit, correspondingly.

In such cases, KCSAN may be safely silenced with data_race(). But in most
cases where KCSAN detects concurrency we should carefully review the code
and likely protect critical sections or switch to atomic find_and_bit(),
as appropriate.

This patch adds the following atomic primitives:

	find_and_set_bit(addr, nbits);
	find_and_set_next_bit(addr, nbits, start);
	...

Here find_and_{set,clear} part refers to the corresponding
test_and_{set,clear}_bit function. Suffixes like _wrap or _lock
derive their semantics from corresponding find() or test() functions.

For brevity, the naming omits the fact that we search for zero bit in
find_and_set, and correspondingly search for set bit in find_and_clear
functions.

The patch also adds iterators with atomic semantics, like
for_each_test_and_set_bit(). Here, the naming rule is to simply prefix
corresponding atomic operation with 'for_each'.

This series is not aimed on performance, but some performance
implications are considered.

In [1] Jan reported 2% slowdown in a single-thread search test when
switching find_bit() function to treat bitmaps as volatile arrays. On
the other hand, kernel robot in the same thread reported +3.7% to the
performance of will-it-scale.per_thread_ops test.

Assuming that our compilers are sane and generate better code against
properly annotated data, the above discrepancy doesn't look weird. When
running on non-volatile bitmaps, plain find_bit() outperforms atomic
find_and_bit(), and vice-versa.

So, all users of find_bit() API, where heavy concurrency is expected,
are encouraged to switch to atomic find_and_bit() as appropriate.

The 1st patch of this series adds atomic find_and_bit() API, 2nd adds
a basic test for new API, and all the following patches spread it over
the kernel.

[1] https://lore.kernel.org/lkml/634f5fdf-e236-42cf-be8d-48a581c21660@alu.unizg.hr/T/#m3e7341eb3571753f3acf8fe166f3fb5b2c12e615

---
v1: https://lore.kernel.org/netdev/20231118155105.25678-29-yury.norov@gmail.com/T/
v2: https://lore.kernel.org/all/20231204185101.ddmkvsr2xxsmoh2u@quack3/T/
v3: https://lore.kernel.org/linux-pci/ZX4bIisLzpW8c4WM@yury-ThinkPad/T/
v4:
 - drop patch v3-24: not needed after null_blk refactoring;
 - add patch 34: "MIPS: sgi-ip27: optimize alloc_level()";
 - add patch 35: "uprobes: optimize xol_take_insn_slot()";
 - add patches 36-40: get rid of locking scheme around bitmaps;
 - move new API to separate headers, to not bloat bitmap.h @ Linus;
 - patch #1: adjust comments to allow returning >= @size;
 - rebase the series on top of current master.

Yury Norov (40):
  lib/find: add atomic find_bit() primitives
  lib/find: add test for atomic find_bit() ops
  lib/sbitmap; optimize __sbitmap_get_word() by using find_and_set_bit()
  watch_queue: optimize post_one_notification() by using
    find_and_clear_bit()
  sched: add cpumask_find_and_set() and use it in __mm_cid_get()
  mips: sgi-ip30: optimize heart_alloc_int() by using find_and_set_bit()
  sparc: optimize alloc_msi() by using find_and_set_bit()
  perf/arm: use atomic find_bit() API
  drivers/perf: optimize ali_drw_get_counter_idx() by using
    find_and_set_bit()
  dmaengine: idxd: optimize perfmon_assign_event()
  ath10k: optimize ath10k_snoc_napi_poll()
  wifi: rtw88: optimize the driver by using atomic iterator
  KVM: x86: hyper-v: optimize and cleanup kvm_hv_process_stimers()
  PCI: hv: Optimize hv_get_dom_num() by using find_and_set_bit()
  scsi: core: optimize scsi_evt_emit() by using an atomic iterator
  scsi: mpi3mr: optimize the driver by using find_and_set_bit()
  scsi: qedi: optimize qedi_get_task_idx() by using find_and_set_bit()
  powerpc: optimize arch code by using atomic find_bit() API
  iommu: optimize subsystem by using atomic find_bit() API
  media: radio-shark: optimize the driver by using atomic find_bit() API
  sfc: optimize the driver by using atomic find_bit() API
  tty: nozomi: optimize interrupt_handler()
  usb: cdc-acm: optimize acm_softint()
  RDMA/rtrs: optimize __rtrs_get_permit() by using
    find_and_set_bit_lock()
  mISDN: optimize get_free_devid()
  media: em28xx: cx231xx: optimize drivers by using find_and_set_bit()
  ethernet: rocker: optimize ofdpa_port_internal_vlan_id_get()
  bluetooth: optimize cmtp_alloc_block_id()
  net: smc: optimize smc_wr_tx_get_free_slot_index()
  ALSA: use atomic find_bit() functions where applicable
  m68k: optimize get_mmu_context()
  microblaze: optimize get_mmu_context()
  sh: mach-x3proto: optimize ilsel_enable()
  MIPS: sgi-ip27: optimize alloc_level()
  uprobes: optimize xol_take_insn_slot()
  scsi: sr: drop locking around SR index bitmap
  KVM: PPC: Book3s HV: drop locking around kvmppc_uvmem_bitmap
  wifi: mac80211: drop locking around ntp_fltr_bmap
  mailbox: bcm-flexrm: simplify locking scheme
  powerpc/xive: drop locking around IRQ map

 MAINTAINERS                                  |   2 +
 arch/m68k/include/asm/mmu_context.h          |  12 +-
 arch/microblaze/include/asm/mmu_context_mm.h |  12 +-
 arch/mips/sgi-ip27/ip27-irq.c                |  13 +-
 arch/mips/sgi-ip30/ip30-irq.c                |  13 +-
 arch/powerpc/kvm/book3s_hv_uvmem.c           |  33 +-
 arch/powerpc/mm/book3s32/mmu_context.c       |  11 +-
 arch/powerpc/platforms/pasemi/dma_lib.c      |  46 +--
 arch/powerpc/platforms/powernv/pci-sriov.c   |  13 +-
 arch/powerpc/sysdev/xive/spapr.c             |  34 +-
 arch/sh/boards/mach-x3proto/ilsel.c          |   5 +-
 arch/sparc/kernel/pci_msi.c                  |  10 +-
 arch/x86/kvm/hyperv.c                        |  41 +--
 drivers/dma/idxd/perfmon.c                   |   9 +-
 drivers/infiniband/ulp/rtrs/rtrs-clt.c       |  16 +-
 drivers/iommu/arm/arm-smmu/arm-smmu.h        |  11 +-
 drivers/iommu/msm_iommu.c                    |  19 +-
 drivers/isdn/mISDN/core.c                    |  10 +-
 drivers/mailbox/bcm-flexrm-mailbox.c         |  21 +-
 drivers/media/radio/radio-shark.c            |   6 +-
 drivers/media/radio/radio-shark2.c           |   6 +-
 drivers/media/usb/cx231xx/cx231xx-cards.c    |  17 +-
 drivers/media/usb/em28xx/em28xx-cards.c      |  38 +--
 drivers/net/ethernet/broadcom/bnxt/bnxt.c    |  18 +-
 drivers/net/ethernet/rocker/rocker_ofdpa.c   |  12 +-
 drivers/net/ethernet/sfc/rx_common.c         |   5 +-
 drivers/net/ethernet/sfc/siena/rx_common.c   |   5 +-
 drivers/net/ethernet/sfc/siena/siena_sriov.c |  15 +-
 drivers/net/wireless/ath/ath10k/snoc.c       |  10 +-
 drivers/net/wireless/realtek/rtw88/pci.c     |   6 +-
 drivers/net/wireless/realtek/rtw89/pci.c     |   6 +-
 drivers/pci/controller/pci-hyperv.c          |   8 +-
 drivers/perf/alibaba_uncore_drw_pmu.c        |  11 +-
 drivers/perf/arm-cci.c                       |  25 +-
 drivers/perf/arm-ccn.c                       |  11 +-
 drivers/perf/arm_dmc620_pmu.c                |  10 +-
 drivers/perf/arm_pmuv3.c                     |   9 +-
 drivers/scsi/mpi3mr/mpi3mr_os.c              |  22 +-
 drivers/scsi/qedi/qedi_main.c                |  10 +-
 drivers/scsi/scsi_lib.c                      |   8 +-
 drivers/scsi/sr.c                            |  15 +-
 drivers/tty/nozomi.c                         |   6 +-
 drivers/usb/class/cdc-acm.c                  |   6 +-
 include/linux/cpumask_atomic.h               |  20 ++
 include/linux/find.h                         |   4 -
 include/linux/find_atomic.h                  | 324 +++++++++++++++++++
 kernel/events/uprobes.c                      |  15 +-
 kernel/sched/sched.h                         |  15 +-
 kernel/watch_queue.c                         |   7 +-
 lib/find_bit.c                               |  86 +++++
 lib/sbitmap.c                                |  47 +--
 lib/test_bitmap.c                            |  62 ++++
 net/bluetooth/cmtp/core.c                    |  11 +-
 net/smc/smc_wr.c                             |  11 +-
 sound/pci/hda/hda_codec.c                    |   8 +-
 sound/usb/caiaq/audio.c                      |  14 +-
 56 files changed, 747 insertions(+), 493 deletions(-)
 create mode 100644 include/linux/cpumask_atomic.h
 create mode 100644 include/linux/find_atomic.h

Comments

Linus Torvalds June 20, 2024, 6 p.m. UTC | #1
On Thu, 20 Jun 2024 at 10:57, Yury Norov <yury.norov@gmail.com> wrote:
>
>
> The typical lock-protected bit allocation may look like this:

If it looks like this, then nobody cares. Clearly the user in question
never actually cared about performance, and you SHOULD NOT then say
"let's optimize this that nobody cares about":.

Yury, I spend an inordinate amount of time just double-checking your
patches. I ended up having to basically undo one of them just days
ago.

New rule: before you send some optimization, you need to have NUMBERS.

Some kind of "look, this code is visible in profiles, so we actually care".

Because without numbers, I'm just not going to pull anything from you.
These insane inlines for things that don't matter need to stop.

And if they *DO* matter, you need to show that they matter.

               Linus
Yury Norov June 20, 2024, 6:32 p.m. UTC | #2
On Thu, Jun 20, 2024 at 11:00:38AM -0700, Linus Torvalds wrote:
> On Thu, 20 Jun 2024 at 10:57, Yury Norov <yury.norov@gmail.com> wrote:
> >
> >
> > The typical lock-protected bit allocation may look like this:
> 
> If it looks like this, then nobody cares. Clearly the user in question
> never actually cared about performance, and you SHOULD NOT then say
> "let's optimize this that nobody cares about":.
> 
> Yury, I spend an inordinate amount of time just double-checking your
> patches. I ended up having to basically undo one of them just days
> ago.

Is that in master already? I didn't get any email, and I can't find
anything related in the master branch.

> New rule: before you send some optimization, you need to have NUMBERS.

I tried to underline that it's not a performance optimization at my
best. People notice some performance differences, but it's ~3%, no
more.

> Some kind of "look, this code is visible in profiles, so we actually care".

The original motivation comes from a KCSAN report, so it's already
visible in profiles. See [1] in cover letter. This series doesn't fix
that particular issue, but it adds tooling that allow people to search
and acquire bits in bitmaps without firing KCSAN warnings.

This series fixes one real bug in the codebase - see #33, and
simplifies bitmaps usage in many other places. Many people like
it, and acked the patches.

Again, this is NOT a performance series.

Thanks,
Yury

> Because without numbers, I'm just not going to pull anything from you.
> These insane inlines for things that don't matter need to stop.
> 
> And if they *DO* matter, you need to show that they matter.
> 
>                Linus
Linus Torvalds June 20, 2024, 7:26 p.m. UTC | #3
On Thu, 20 Jun 2024 at 11:32, Yury Norov <yury.norov@gmail.com> wrote:
>
> Is that in master already? I didn't get any email, and I can't find
> anything related in the master branch.

It's 5d272dd1b343 ("cpumask: limit FORCE_NR_CPUS to just the UP case").

> > New rule: before you send some optimization, you need to have NUMBERS.
>
> I tried to underline that it's not a performance optimization at my
> best.

If it's not about performance, then it damn well shouldn't be 90%
inline functions in a header file.

If it's a helper function, it needs to be a real function elsewhere. Not this:

 include/linux/find_atomic.h                  | 324 +++++++++++++++++++

because either performance really matters, in which case you need to
show profiles, or performance doesn't matter, in which case it damn
well shouldn't have special cases for small bitsets that double the
size of the code.

              Linus
Yury Norov June 20, 2024, 8:20 p.m. UTC | #4
On Thu, Jun 20, 2024 at 12:26:18PM -0700, Linus Torvalds wrote:
> On Thu, 20 Jun 2024 at 11:32, Yury Norov <yury.norov@gmail.com> wrote:
> >
> > Is that in master already? I didn't get any email, and I can't find
> > anything related in the master branch.
> 
> It's 5d272dd1b343 ("cpumask: limit FORCE_NR_CPUS to just the UP case").

FORCE_NR_CPUS helped to generate a better code for me back then. I'll
check again against the current kernel.

The 5d272dd1b343 is wrong. Limiting FORCE_NR_CPUS to UP case makes no
sense because in UP case nr_cpu_ids is already a compile-time macro:

#if (NR_CPUS == 1) || defined(CONFIG_FORCE_NR_CPUS)
#define nr_cpu_ids ((unsigned int)NR_CPUS)
#else
extern unsigned int nr_cpu_ids;
#endif

I use FORCE_NR_CPUS for my Rpi. (used, until I burnt it)

> > > New rule: before you send some optimization, you need to have NUMBERS.
> >
> > I tried to underline that it's not a performance optimization at my
> > best.
> 
> If it's not about performance, then it damn well shouldn't be 90%
> inline functions in a header file.
> 
> If it's a helper function, it needs to be a real function elsewhere. Not this:
> 
>  include/linux/find_atomic.h                  | 324 +++++++++++++++++++
> 
> because either performance really matters, in which case you need to
> show profiles, or performance doesn't matter, in which case it damn
> well shouldn't have special cases for small bitsets that double the
> size of the code.

This small_const_nbits() thing is a compile-time optimization for a
single-word bitmap with a compile-time length.

If the bitmap is longer, or nbits is not known at compile time, the
inline part goes away entirely at compile time.

In the other case, outline part goes away. So those converting from
find_bit() + test_and_set_bit() will see no new outline function
calls.

This inline + outline implementation is traditional for bitmaps, and
for some people it's important. For example, Sean Christopherson
explicitly asked to add a notice that converting to the new API will
still generate inline code. See patch #13.
Linus Torvalds June 20, 2024, 8:32 p.m. UTC | #5
On Thu, 20 Jun 2024 at 13:20, Yury Norov <yury.norov@gmail.com> wrote:
>
> FORCE_NR_CPUS helped to generate a better code for me back then. I'll
> check again against the current kernel.

Of _course_ it generates better code.

But when "better code" is a source of bugs, and isn't actually useful
in general, it's not better, is it.

> The 5d272dd1b343 is wrong. Limiting FORCE_NR_CPUS to UP case makes no
> sense because in UP case nr_cpu_ids is already a compile-time macro:

Yury, I'm very aware. That was obviously intentional. the whole point
of the commit is to just disable the the whole thing as useless and
problematic.

I could have just ripped it out entirely. I ended up doing a one-liner instead.

                Linus